M-PACT: AI-Driven Methylation Liquid Biopsy for Pediatric CNS Tumor Classification

Pediatric brain tumors remain the leading cause of cancer-related death in children, in part because they represent not one disease but a sprawling collection of biologically distinct malignancies. Over the last decade, molecular profiling—especially DNA methylation classification—has reshaped how these tumors are defined and diagnosed, offering a kind of epigenetic fingerprint that can separate entities that look similar under a microscope but behave very differently in the clinic.

The problem is that the modern diagnostic toolbox still depends heavily on tissue. Some tumors sit in high-risk, surgically inaccessible locations. Others yield scant biopsy material. And repeated sampling to track evolution or recurrence is rarely feasible—particularly for metastatic disease.

That reality has made liquid biopsy an alluring alternative. In many adult cancers, blood-based assays transformed care by detecting tumor-derived DNA and guiding therapy. Central nervous system tumors are a tougher target. The blood–brain barrier limits how much tumor material reaches plasma, reducing sensitivity. Cerebrospinal fluid (CSF), by contrast, has emerged as a richer compartment for tumor-derived cell-free DNA (cfDNA), collected intraoperatively or through lumbar puncture or ventricular access. Prior studies have shown that CSF can support risk stratification and monitoring using approaches that range from targeted mutation tests to genome-wide methods such as low-coverage whole-genome sequencing (lcWGS) or Nanopore sequencing. But the field has struggled with a recurring set of constraints: low cfDNA yields (often subnanogram), low tumor fractions, and a reliance on genetic alterations—mutations or copy-number variations (CNVs)—that may be absent or hard to detect in many pediatric cases.

The work described here proposes a different anchor point: methylation. Rather than treating genetic alterations as the primary biomarkers, the investigators developed M-PACT (methylation-based predictive algorithm for CNS tumors), a deep neural network designed to classify pediatric CNS tumors from CSF-derived cfDNA methylomes even when inputs are extremely small. The pipeline is built around three key strategies: imputing missing CpG methylation values to compensate for sparse sequencing, enriching for tumor signal by subtracting predicted normal background, and using methylation-based deconvolution to estimate malignant and nonmalignant fractions once a tumor classification is achieved.

The team first stress-tested the concept computationally. Using a reference cohort of 914 CNS tumor methylation arrays, they trained a network-based regression diffusion model that could accurately impute methylation values even when only 100,000 CpGs were observed. They then generated 3,195,000 in silico samples by varying tumor fraction and CpG sparsity across 84 tumor and 13 non-tumor entities, training three neural networks—one generic and two tuned for sparse, low–tumor fraction conditions—and combining them into a layered ensemble. In this simulated setting, M-PACT achieved strong performance: average F1 scores of 0.95 when tumor fraction exceeded 0.5 and 0.91 when it was below 0.5, with performance predictably degrading when tumor fraction dropped further and CpG sparsity increased.

A classifier, however, is only as useful as the real-world assay that feeds it. To generate methylomes from the low cfDNA inputs typical of pediatric CSF samples, the investigators optimized enzymatic methylation sequencing (EM-seq) for cfDNA. Across 268 cfDNA samples, median input was 0.5 ng (with some as low as 0.05 ng), producing a median of 7 million CpG sites captured at roughly 4× genome-wide coverage. Importantly, EM-seq was evaluated not only as a methylation platform but also for CNV detection—a practical bridge to first-generation CSF assays. In 110 samples with matched lcWGS and EM-seq, genome-wide CNV profiles were highly concordant, with a median scaled Manhattan similarity of 0.94 and a strong correlation in CNV-based ctDNA fraction estimates (Pearson r = 0.928), suggesting EM-seq can carry forward CNV-based capabilities while enabling methylation classification.

From there, the clinical-relevant question becomes classification accuracy in patient CSF. In an embryonal CNS tumor benchmarking cohort restricted to CNV-positive CSF samples (79 participants), M-PACT matched tissue-based methylation classifications in 73 of 79 cases (92%). In an independent embryonal validation cohort (58 CSF samples from 48 participants), accuracy was 51 of 58 (88%). The pattern behind failures was consistent: incorrectly classified or unclassifiable samples had significantly lower ctDNA fractions. CpG sparsity, by contrast, was not the main driver in the embryonal cohorts. When the investigators broadened the test to nonembryonal tumors—including high-grade glioma, ependymoma, germ cell tumor, choroid plexus tumor, meningioma, and BCOR-altered high-grade neuroepithelial tumor—M-PACT correctly classified 22 of 29 samples (76%), with misclassifications now associated with both lower ctDNA fraction and higher CpG sparsity. Across cohorts, a ctDNA fraction threshold of 0.15 corresponded to 95% classification accuracy.

One of the more consequential demonstrations comes from the cases where CNVs are not available as a crutch. Some pediatric CNS tumors have “balanced” genomes without discernable CNVs, limiting the utility of CNV-dependent liquid biopsy strategies. In a small series of participants with balanced tumor genomes, M-PACT still classified CNV-negative cfDNA profiles with high probability (0.85–0.98), reporting methylation-based ctDNA fractions ranging from 3% to 68%. In the monitoring setting, longitudinal CSF sampling in one medulloblastoma case showed that M-PACT classifications tracked the disease course, illustrating how methylation patterns could serve as a dynamic readout even when CNVs are uninformative.

The study also pushes beyond “what tumor is this?” toward “what else can the methylome tell us?” By applying deconvolution using reference signatures from nine nonmalignant cell classes, the pipeline estimated the relative contributions of immune and neural lineages within nonmalignant CSF and within tumor-associated CSF samples. In medulloblastoma subgroups, inferred differences appeared in the proportions of oligodendrocytes, B cells, and astrocytes—an example of how CSF methylomes could potentially reflect shifts in the CNS microenvironment alongside tumor burden.

Clinically, the vignettes are designed to mirror dilemmas that neuro-oncologists face. Intraoperative CSF collected at diagnosis yielded high ctDNA fractions (57–88% by methylation deconvolution) and classifications concordant with tumor tissue, suggesting a route toward molecular diagnosis even when tissue is limited or surgery is risky. In follow-up, CSF profiling helped discriminate suspected recurrence from secondary malignancy: in two participants previously treated for medulloblastoma, EM-seq detected CNV hallmarks such as PDGFRA amplification and CDKN2A/B deletion and M-PACT classified the CSF profiles as glioblastoma, aligning with the concept of radiation-induced secondary glioma and highlighting how a CSF-based molecular diagnosis could redirect management when imaging is ambiguous.

The authors are candid about remaining limitations. The EM-seq/M-PACT approach does not detect gene-level mutations, a capability still better served by targeted sequencing. And while methylation-based and CNV-based detection achieved thresholds in the range of ~5–10% tumor burden—similar to what many sequencing-based CSF studies have used—the study underscores how much technical innovation is still needed to push sensitivity lower in the low-yield CSF environment. A complementary “binary” methylation algorithm, comparing CSF methylomes to participant-matched tumor tissue and to control CSF/brain references, improved ctDNA detection to 96% among CNV-positive samples, rescuing some cases below M-PACT’s classification threshold.

Taken together, this work presents a blueprint for bringing methylation-first thinking into CSF liquid biopsy for pediatric CNS tumors: a workflow that tolerates subnanogram inputs, classifies across tumor entities with strong concordance to tissue standards, retains CNV detection comparable to lcWGS, and adds the ability to infer cellular composition in the CSF compartment. The authors position the next step clearly: prospective validation to determine how this pipeline performs when deployed in real-time clinical decision-making, where the highest value may lie not just in confirming diagnoses, but in enabling earlier, less invasive, and more informative molecular tracking of pediatric brain tumors over time.